Will machines ever replace man when it comes to completing normal everyday tasks? While we debate and ponder this possibility, there is a research lab at the University of Toronto that’s making huge strides in the artificial intelligence field with their research in machine learning.
Last June, two U of T researchers, Sanja Fidler and Raquel Urtasun, developed an algorithm that analyzes your clothes and gives you feedback on how to improve your personal style.
The algorithm uses “deep learning,” a subset of machine learning technologies. This sounds like something pulled directly from a sci-fi movie; when in fact, U of T is home to one of the leading research labs specializing in machine learning. Furthermore, we’re more accustomed to uses of artificial intelligence in our every day lives through applications such as Apple’s Siri.
Using image posts from the website Chictopia, a fashion blogging website, the research team, which also includes two researchers from Barcelona’s Institute of Robotics and Industrial Informatics (IRI), analyzed over 144,000 posts over the course of a year. Chictopia allows users to publish posts of their personal style, using a tagging system to reach out and connect with other users who are similar in age, shape, style, and other characteristics.
The algorithm is capable of learning and thinking through the intake of data from the Chictopia website. It uses this data to determine the current trends for different geographic groups. Then it teaches itself to identify new fashion trends as more Chictopia posts are made.
“The algorithm is ‘taught’ on [Chictopia], a site where style enthusiasts post their latest photos, possibly tag them (with a list of certain clothing items they are wearing), and then other users comment on the post and also have the option to ‘like’ or vote for the post,” Fidler said.
“The algorithm looks at all this data… and tries to predict how many votes this photo would get. The number of votes is our proxy for ‘fashionability,’” Fidler further explained.
On the user-end, one would give the algorithm a photo to be analyzed, the algorithm would then examine your clothes, as well as the finer details of the image such as your facial expression, body type, composition of the photo, as well as scenery. With the extracted information, it will then think about whether or not the outfit fed to them was the best possible outfit the user could have selected. In the end, it spits out suggestions to improve one’s style which may include a change of skirt or a colour of shoes.
“Our app will not be primarily designed for a designer but for regular people (like us!),” Fidler said, “[f]or example, you are going to a party and you want to look at your best. You take a photo of yourself and the app will tell you whether you look stylish or not (for the particular event) and what you could change to improve your look.”
The fashion algorithm is just one of the many projects that are part of research that looks at different ways in which deep learning can be adapted to everyday use. Although most of the fashion algorithm has been used for research purposes, there is currently a team working on developing an app for mainstream use.
Together with the researchers from IRI, Fidler and Urtasun presented their findings in a research paper, “Neuroaesthetics in Fashion: Modeling the Perception of Fashionability” at the 2015 Conference on Computer Vision and Pattern Recognition (CVPR) in June.
